Proteomic Pattern Classification Using Bio-markers for Prostate Cancer Diagnosis

Abstract

Decision trees (DTs) and multi-layer perceptron (MLP) neural networks have long been successfully used to various pattern classification problems. Those two classification models have been applied to a number of diverse areas for the identification of ‘biologically relevant’ molecules. Surface enhanced laser desorption/ionization time-of-flight mass spectrometry (SELDITOF MS) is a novel approach to biomarker discovery and has been successfully used in projects ranging from rapid identification of potential maker proteins to segregation of abnormal cases from normal cases. SELDI-TOF MS can contain thousands of data points. This high dimensional data causes a more complex neural network architecture and slow training procedure. In the approach we proposed in this paper, a decision tree is first applied to select possible biomarker candidates from the SELDI-TOF MS data. At this stage, the decision tree selects a small number of discriminatory biomarker proteins. This candidate mass data defined by the peak amplitude values is then provided as input patterns to the MLP neural network which is trained to classify the mass spectrometry patterns. The key feature of this hybrid approach is to take advantage of both models: use the neural network for classification with significantly lowdimensional mass data obtained by the decision tree. We applied this bioinformatics tool to identify proteomic patterns in serum that distinguish prostate cancer samples from normal or benign ones. The results indicate that the proposed method for mass spectrometry analysis is a promising approach to classify the proteomic patterns and is applicable for the significant clinical diagnosis and prognosis in the fields of cancer biology.